Lustre usages and experiences

Transcription

1 Lustre usages and experiences at German Climate Computing Centre in Hamburg Carsten Beyer

2 High Performance Computing Center Exclusively for the German Climate Research Limited Company, non-profit Staff: ~ 70 Services for Climate Research: Support for Scientific Computing and Simulation, Model Optimization, Parallelization Data Management and Archiving Data Visualization (3D Graphics and Video) About DKRZ University Research Group: HPC (Prof. Dr. Ludwig) 2 / 18

3 Mistral First phase 2015 (second phase 2016), total cost: 41 Mio Euro Bull Supercomputer: 26 Mio Euro Bullx B700 DLC-System ~ (+ ~67.000) cores (Intel Haswell/Intel Broadwell) nodes (2x 12 Cores) (+1750 nodes 2x 18 Cores) 1,4 (3,0) PetaFLOPS 115 TB (266 TB) main memory Infiniband FDR Parallel file system: Lustre, 21 (+33) PetaByte Throughput > 0.5 TeraByte/s 3 / 18

4 Lustre - ClusterStor Seagate ClusterStor Setup (Phase 1 - CS9000/ Phase 2 - L300) 62 OSS / 124 OST s / 6TB disks 5 MDT 21 PB / max. 6 Billion files Lustre / IB FDR 455 million files / 90% filled 74 OSS / 148 OST s / 8TB disks 7 MDT 33 PB / max. 8 Billion files Lustre / IB FDR 196 million files / 40% filled 4 / 18

5 Filesystem structure Lustre Phase 1 HOME directories (/mnt/lustre01/pf => MDT0000) POOL directory (/mnt/lustre01/pool => MDT0000) Software tree (/mnt/lustre01/sw => MDT0000) SCRATCH dirs (/mnt/lustre01/scratch => MDT000[1-4]) WORK-directories (/mnt/lustre01/work => MDT000[1-4]) Lustre Phase 2 WORK directories (/mnt/lustre02/work => MDT000[0-6]) extension of phase 1 Soft link /work -> /mnt/lustre01/work Soft link /mnt/lustre01/work/<prj> -> /mnt/lustre02/work/<prj> 5 / 18

6 Migration GPFS to Lustre Copying 4.5 PB from GPFS (AIX) to Lustre (Linux) Usage of GPFS policies to generate filelist (130 mio files) Sort and split filelist (mix of big/medium/small files) as input for rsync Using SLURM on new system to schedule ~6000 Jobs for copying via IB gateway from NSD server Challenge Not overloading NSD server on previous system (2x 10 GbE per node) Changing UID/GID for some users from id s <1000 to >20000 during transfer (newer rsync version needed than in RHEL 6) Time frame to achieve this Because Benchmarks for approval of the system still ongoing 6 / 18

7 Migration GPFS to Lustre Where s my Quota and why takes mv so long All copied project data belonging to MDT0000 in a separate location The new project directories were distributed to MDT000[1-4] How to copy the old data to the new directories? Could not be moved easily with mv due to DNE phase 1 Trying tools like shift for copying No quota could be used as before (user/fileset/user in fileset) HOME/SCRATCH/WORK in one filesystem now Using Robinhood for soft quota (starting 2016) User can see their quota and amount of data on a DKRZ web frontend 7 / 18

8 Migration GPFS to Lustre Small files HOME directories (~ 30 mio files / 6 TB ) Software tree Problem to run Backup with Calypso/Simpana Full Backup of HOME takes days Large loading times for software (e.g. Matlab, Python) Generating a 300GB image file stripped accross 16 OST s Loop mount of this image file to interactive nodes (Login/Graphics) Using caching on clients for higher performance 8 / 18

9 Tools used - Copytools Self-Healing Independent File Transfer (Shift) Paul Kolano (NASA) Dependent on mutil (Paul Kolano / NASA) Final rsync at the needed (Hardlinks) lfs find + rsync Generating filelist with lfs find and split in equal parts rsync with split-lists (running as SLURM jobs in parallel) Final rsync needed afterwards for hardlinks, directory ownership pftool Los Alamos National Lab Easy to use, scalable on several nodes with SLURM Needs final rsync afterwards for hardlinks, directory ownership 9 / 18

10 Tools used - Robinhood Report generation for User reports -> HOME/SCRATCH Project reports -> WORK Total amount of data for each project (~ 220 active projects) Per user in a project (up to 130 user per project) Overview for project admin Soft quota by mail to users / project admin Setup Robinhood version / reading Lustre changelogs 2 Robinhood server for lustre01 (connected to 2/3 MDT) 3 Robinhood server for lustre02 (connected to 2/2/3 MDT) 10 / 18

11 Tools used HPSS pftp Tape Archive HPSS Projects have to request storage on a annual basis (Quota) Manual copy of data by user for their projects with pftp User can decide for single / double copy or long term archive (LTA) Douple copy is also accounted twice for project LTA needs a description of data (to be identified later, stored up to 10 years) Currently no automatic migration from Lustre to HPSS Possible tool could be Robinhood Currently not tried yet How to identify data to be migrated for projects 11 / 18

12 Monitoring tools / sources Sources: Starting with own python daemon on CS (replaced by Seastream) Seastream API on ClusterStor (currently switched off on lustre02) Entries from syslog (Lustre clients) Lustre lllite (from clients) Tools: Opentsdb / hbase / hadoop / elastic search Frontends: Xdmod / Grafana / Kibana Icinga on Clusterstor Thanks to: Olaf Gellert, Hendrik Bockelmann, Josef Dvoracek (ATOS), Eugen Betke (Uni HH) 12 / 18

13 Monitoring - Overview 3340 CN cluster 24 Login On demand: System metrics Lustre llite Mistral Permanent: System metrics Lustre llite Lustre Server FS lustre01 Seastream Job stats System stats Lustre Server FS lustre02 Nginx proxy Syslog-ng cluster SLURMctld Nightly testsuite with IOR, Climate models Opentsdb 6 instances hbase 6 instances HADOOP 4 instances Log files Nginx proxy Elastic search 4 instanses Job scripts xdmod grafana kibana 13 / 18

14 Monitoring - Grafana 14 / 18

15 Monitoring - Grafana 15 / 18

16 Pro s / Con s from daily business Most times fine but Degradation of performance when reaching >95% filling state and not symmetric usage of OST s (lustre01) First solution was to deactivate OST S / migrating data from OST s and shifting data to lustre02 filesystem Newer solution is setting qos_threshold_rr to 5% on all MDT That was leading to about 1.2 PB of unusable disk space With the newer setting it is below 1 PB Degradation of performance by monthy RAID-check (lustre01) Lowering the raid check priority RAID check runs now 2 ½ weeks 16 / 18

17 Pro s / Con s from daily business Client dis-/reconnects from OST s / MDT s Issue for SLURM jobs with IO (extended runtime) Robinhood could not read changelog anymore Workaround: Failover/reboot/failback affected MDT Firmware bug on hardware (lustre02) Rare case of watchdog issue causes OSS HA pair to power down After power up of OSS and fixing OST s reconnect from clients or umount of lustre02 not possible Reboot of all Lustre clients needed (power reset) 17 / 18

18 Thank you for your Attention! Carsten Beyer Questions? 18 / 18